Histopathology is a cornerstone of cancer diagnosis, but standard HE-stained images show tissue morphology, not the underlying molecular activity that drives disease behavior. The new preprint introduces REST, a method that aims to predict gene expression directly from these images.
REST uses spatial transcriptomics data as supervision during training, providing a molecular ground truth that is aligned with tissue structure. The "reliability-guided" component suggests the model accounts for the varying quality or trustworthiness of that supervision signal, though the abstract does not give implementation details.
If effective, such an approach could let researchers and clinicians extract molecular information from routine pathology slides without additional staining or sequencing. The preprint is available on arXiv.